Machine- and deep-learning models for wind turbine icing prediction across multiple horizons: the influence of ice sensors and weather forecasts

Around one in four onshore wind turbines operates in cold-climate regions – representing more than three times the global capacity of offshore wind – where annual energy production can be significantly reduced due to ice accretion. Despite growing interest in machine learning for turbine-icing prediction, previous studies have been mostly limited to SCADA data and short-term forecasts, offering little value for operational decision-making. We examine how newly available icing-sensor measurements and liquid water content (LWC) forecasts, together with AI model design, expand what is now possible in turbine-icing prediction. Using one winter (December 2023 to April 2024) of high-resolution data from a single cold-climate wind turbine, including SCADA, on-site meteorological data, icing measurements and weather forecast variables, five models were benchmarked: three ensemble learners (Random Forest, XGBoost, LightGBM) and two deep-learning architectures (CNN-LSTM, GRU) across three prediction horizons (1 h, 6 h, 24 h) and four input-richness configurations. For the studied site and winter season, the results reveal a clear pattern: physically meaningful inputs, rather than network depth, dominate predictive performance. At the 1 h horizon, adding icing indicators and LWC improved mean F1-score from 0.76 to 0.83, reflecting the benefit of direct microphysical information. At 6 h and 24 h, incorporating forecast-based inputs produced the largest gains, increasing mean F1 from 0.46 to 0.70 and from 0.41 to 0.64, respectively. Once supplied with rich physical inputs, lightweight ensemble models matched the accuracy of deep networks, supporting efficient edge-AI deployment directly at turbine level. These findings show that emerging ice-sensor and forecast data expand what is possible in AI-based icing prediction. They create new opportunities for short-horizon icing protection system (IPS) control through earlier intervention, and for more informed day-ahead market participation through better anticipation of icing-related production losses in cold climates.